Losing Control of AI Is Actually the Plan

Summary

In this opinion essay, Kelsey Piper argues that leading AI labs are preparing to hand an increasing share of AI development to AI systems themselves. The article says OpenAI and Anthropic are building internal automated research operations intended to help create more capable successors, with OpenAI saying it is making progress toward an automated AI researcher by March 2028. Piper cites OpenAI’s claim that the share of research compute devoted to internal coding inference rose 100-fold over six months and Anthropic’s statement that most of its work is internal software engineering. AI systems already assist with coding, experiments, reinforcement learning, and other model-development tasks, which the labs say can accelerate progress. The article presents recursive self-improvement as the next step: models would help train and improve their successors while human researchers mainly interpret reports and check whether systems remain on task. Piper argues that the volume and speed of this work could exceed available human attention, making the systems’ behavior difficult to understand or audit. The essay also notes that researchers acknowledge these risks while defending continued progress, often invoking competition with other companies or China. Its conclusion is an argument for stronger external oversight and regulation, rather than a neutral prediction that recursive self-improvement will occur.